Only Nvidia can beat Nvidia.
Last night, Nvidia officially released the next-generation SoC chip – Thor. This is a performance monster with a single-chip computing power of up to 2,000 TOPS, which is nearly eight times the calculation power of the current product Orin, and 28 times that of a single FSD chip.
At last year’s Nvidia GTC conference, Nvidia released a new generation of products – Atlan, with a computing power of 1,000 TOPS. But less than a year later, Nvidia has iterated itself.
While chip suppliers were still developing flagship chips to compete with Nvidia Orin, Nvidia had already come up with the next-generation product and won clients such as Jikr.
In the name of “Thor”
Nvidia’s naming of its chip products is also interesting. The current product Orin is named after the son of Atlantis mythology’s ruler of the sea, the sea king Atlan. According to Nvidia’s original plan, they would launch the Atlan autonomous driving chip with greater computing power and stronger overall performance in 2025. However, due to the dominance of the next-generation product and the nearly identical mass production time, Atlan was ultimately stillborn.
This new flagship that is powerful enough to “beat” Nvidia’s own products does not continue the naming tradition of “ocean mythology”, but is named “Thor” (Thunder God Thor).
Thor has the following characteristics: high performance, high integration, and support for large models.
The strongest SoC
If you list Nvidia’s chip product parameters in the field of intelligent automobiles in the past few years, you will find that their product development has been “leapfrogging”. When their peers were still developing large-scale computing power chips of several hundred TOPS, Nvidia had already determined that they could mass-produce chips with 2,000 TOPS in 2025.
Thor integrates 77 billion transistors, but Nvidia has not disclosed Thor’s process technology.Lao Huang said that the performance indicators of Thor are achieved by upgrading the CPU (Grace), GPU (Ada Lovelace), and the engine for processing Transformer models (Hopper).
Grace is a highly specialized processor designed for large-scale data-intensive HPC (high-performance computing) and AI applications. Lao Huang said that all previous parallel algorithms have been accelerated by GPUs, and other workloads are often limited by single-threaded performance, and Grace has excellent single-threaded performance, which can make up for it.
Hopper is an engine for processing Transformer models, which provides amazing fast transformation capabilities for Transformer and Vision Transformer. Ada Lovelace is the newly released architecture for the 40 series graphics card, using 4nm process, with a shader capability of up to 83 TFLOPS for streaming multi-processors, and throughput is twice that of the previous generation product.
Not only the automatic driving chip
In the past, traditional cars may have installed hundreds of ECUs, whose function is to control different functions. Increasing vehicle functionality means the need to increase corresponding control modules and wire harnesses.
The total length of wire harnesses on a traditional car can reach several kilometers, and the total weight can also reach dozens of kilograms, which is costly and difficult to control. Nowadays, smart cars have evolved from distributed to domain controller architecture.
The industry’s consensus is that centralized central domain controllers replace distributed domain controllers, namely, a central brain controls all vehicle functions.
Currently, the solutions of Chinese new forces and mainstream OEMs generally adopt Nvidia or other supplier chips for intelligent driving domains, while Qualcomm solutions are adopted for cabin domain controllers.
To achieve the “unification”, a chip that can span across different fields is needed, and Thor is such a chip.
Thor integrates all AI computing requirements in the smart car field, including intelligent driving, active safety, smart cockpit, automatic parking, in-car operating systems, and entertainment, etc.
Most importantly, Thor can be configured into multiple modes, and its 2,000 TOPS computing power can be flexibly allocated. Users can call the massive computing power to different tasks according to their needs.# Thor’s Multi-Source Domain Isolation Enables Concurrent and Time-Sensitive Process Execution Without Interruption
On a single computer, vehicles can simultaneously run Linux, QNX, and Android with Thor’s multi-source domain isolation, which allows for concurrent and time-sensitive process execution without interruption.
It is also worth noting that like Orin, Thor is not limited to the smart automotive industry and is also applicable to multiple areas, such as healthcare and industry.
Supports Large-scale Models
As mentioned earlier, Hopper, an engine for processing Transformer models, is included in the Thor chip.
Transformer, a model first proposed by Google for natural language processing (NLP), has been widely used in the field of autonomous driving in the past two years.
Tesla was the first to introduce Transformer to the autonomous driving industry, and many domestic automakers and suppliers have followed suit.
The parameter size of the Transformer network has been continuously increasing. With the increasing number of vehicle sensors and the improved information quality obtained by sensors (especially cameras), the system must have the ability to quickly process large-scale sensing data.
The main function of Hopper is to optimize models for large-scale data, thereby improving the system’s data processing capabilities.
Based on the development trend of the industry, more and more automakers will adopt Transformer models in their self-developed systems. NVIDIA has seized this opportunity and deployed in a timely manner.
Leading the Charge Again: Jicai’s “New Force-ification”
When Orin was released, many new forces in China rushed to be the first to adopt it. However, with the release of Thor, NVIDIA took the initiative to cue Jicai.
Jicai currently has two products, 001 and 009. Leaving aside the as-yet-unreleased 009, only the 001 is the product that customers can currently get their hands on. In fact, Thor is Jicai’s second time leading the charge as a first-mover supplier of the latest chips, and let’s not forget that Jicai 001 was the world’s first car model powered by Mobileye EyeQ5.
From this perspective, Jicai has always attached great importance to advanced driving assistance systems (ADAS).
However, due to a variety of reasons, the functional landing progress of EyeQ5 in China has been very slow. To put it severely, in today’s “crazy” era of intelligentization, if Jicai puts all its hopes on Mobileye, it will undoubtedly be passive and wait to die. Therefore, Jicai has engaged in self-developed work.### Mobileye and Self-Development Go Hand in Hand
Strictly speaking, Jidu Auto’s intelligent driving business has three main lines. The first is the autonomous driving business in partnership with Waymo. The second is the adoption of Mobileye technology supplied by vendors to provide high-level assisted driving for mass-produced vehicles. The third involves developing Jidu’s own smart driving technology from scratch.
Currently, the same team is working on all three lines, but more attention is being paid to the implementation of smart driving features based on the EyeQ5 chip and self-development.
Why Self-Development?
Whenever the question of “why self-development” is raised to the research and development personnel in the automotive industry, the answer always centers around autonomy.
Jidu Auto has the same concerns, and they are the most significant.
When Jidu Auto released the 001 model, the assisted driving hardware on this car was the source of infinite imagination, especially the industry-forward-looking deployment of the first EyeQ5 and 8 million-pixel vision. It made us believe that the 001 model had the best chance of breaking into the top tier of vehicles. But the development of EyeQ5 chip capabilities has been slower than we imagined domestically, and user grievances have accumulated.
The only way to appease users, while keeping up with the industry, is to reclaim the initiative in research and development while satisfying user needs as much as possible.
A recent Jidu Auto update brought an improvement in LCC functionality, but this is not enough; obviously, possessing only basic assisted driving features cannot meet user needs.
When asked about the progress of Jidu’s research and development based on the Mobileye EyeQ5 chip, Jidu Auto revealed that R&D has been ongoing and that there are already internal testing versions of NZP and AutoPark.
Conclusion
After betting on Mobileye and wasting two years, Jidu Auto must grab the next round of arms races and the trend of self-development.Choosing to develop our own technology or partnering with NVIDIA is a correct but difficult decision in the current stage for the automotive industry. However, Zeekr Hao has taken the initiative now.
Meanwhile, Zeekr Hao is facing some problems that need to be solved. If the development progress of EyeQ5-based functions remains slow in the next 2 to 3 years, how can Zeekr Hao maintain its competitiveness? When will NZP and automatic parking function be available to users? These are issues that users are concerned about.
Since last year, Zeekr Hao has been sending signals that its self-developed technology is gaining momentum. Now, partnering with NVIDIA, Zeekr Hao seems to be more proactive than we have seen.
This article is a translation by ChatGPT of a Chinese report from 42HOW. If you have any questions about it, please email bd@42how.com.